Canopy-258M-R3 v3: Ultra-Accelerated Edge Browser Agent & MoE Flagship

Canopy-258M-R3 v3 is an ultra-efficient 258.56M parameter Recurrent Mixture-of-Experts (MoE) model optimized for high-speed edge computing, tool synthesis, and browser automation.

v3 introduces the Hybrid Input Protocol and Event-Driven DOM Synchronization, achieving a 2.02x speedup across real-world browser batteries and a 5.1x acceleration (500%+) on multi-step interactive forms.

What is New in v3

1. Hybrid Input Protocol (ActionExecutor):

   • Automatically differentiates static form fields (passwords, emails, registration, multi-step wizards) from reactive stream targets (search autocompletes, live comboboxes, typeaheads).
   • Static Form Inputs: Uses atomic CDP fill() with standard DOM input and change event dispatch, dropping typing latency from >1,400 ms down to <40 ms (35x acceleration).
   • Interactive Streams: Automatically detects search/combobox targets and preserves natural Gaussian typing cadence ($\mu=35 ext{ms}$) so reactive debounce and AJAX handlers populate suggestions properly.
   • Developer control supported via BrowserAction(stream_input=True|False).

2. Snappy Cubic Bézier Motor Kinematics:

   • Calibrated cubic Bézier trajectory duration (80 ms – 220 ms, 8–16 steps) with smoothstep ease-in/ease-out curves and sub-pixel micro-jitter.
   • Preserves human motor telemetry to avoid synthetic robotic detection while eliminating sluggish travel lag.

3. Event-Driven Scroll Mutation Synchronization:

   • Streamlined wheel step dispatch with immediate scroll and resize window event notifications.
   • Instantly triggers IntersectionObserver handlers and dynamic infinite scroll listeners without stalling.

4. Structured Data Scraping & Web Extraction Engine:

   • First-class structured parsing for tables (2D row/column matrices), lists, and individual element attributes via controller.scrape_table(), controller.scrape_element(), and controller.scrape_page().

5. Visual Step Logging & Audit Receipts:

   • Step-by-step high-resolution PNG screenshot auditing with illuminated green bounding boxes on target marks, glowing red coordinate reticles, and top-left HUD step badges.

Performance Benchmarks: v2 vs v3 (7-Scenario Battery)

Complex Chained Actions Benchmark (Multi-Stage Orchestration)

Model Architecture Specifications

Quickstart: Python Inference & Web Automation

1. Model Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "psikosen/canopy-258m-r3-v3"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "<|im_start|>user\nWrite a Python function to extract all email addresses from a web page.<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

2. Fast Browser Agent Execution (miniswardbower)

import asyncio
from miniswardbower.browser.controller import BrowserController
from miniswardbower.core.config import BrowserConfig
from miniswardbower.core.schemas import BrowserAction, BrowserActionType

async def run_agent():
    controller = BrowserController(BrowserConfig(headless=True))
    await controller.start()
    try:
        await controller.goto("https://news.ycombinator.com")
        
        # 1. Scrape structured data
        page_data = await controller.scrape_page()
        print("Page Title:", page_data.get("title"))
        
        # 2. Fast atomic fill
        await controller.execute_action(
            BrowserAction(op=BrowserActionType.TYPE, target="input[name='q']", text="LLM Edge", stream_input=False)
        )
        await controller.execute_action(
            BrowserAction(op=BrowserActionType.PRESS, key="Enter")
        )
    finally:
        await controller.stop()

asyncio.run(run_agent())

License

Released under the Apache 2.0 License.